mallocPitch: A higher-level interface to cudaMallocPitch

In duncantl/RCUDA: R Bindings for the CUDA Library for GPU Computing

Description

mallocPitch is similar to cudaMallocPitch for allocating memory for 2-dimensionsal structures, i.e. matrices. Unlike cudaMallocPitch, this function takes the name of the element type and determines the number of bytes of each element. More than this, it returns an object that is similar to what cudaMallocPitch returns, but contains more information. This allows us to directly assign an R object directly to the allocated memory without having to explicitly call cudaMemcpy2D.

One has to keep in mind that R matrices are stored in column order but the Pitch mechanism is more focused on rows in CUDA.

Usage

mallocPitch(width, height, elType = NA)

Arguments

width	the number of columns in the 2D object
height	the number of rows in the 2D object.
elType	the name of the element type of the array, e.g. int, float, etc.

Value

mallocPitch returns a list with elements

devPtr	a reference to the allocated memory on the GPU
pitch	the actual pitch value between the start of elements in consecutive rows.
elType	the value of elType
elSize	the number of bytes for a single element
dim	the dimension vector of the allocated space, i.e. width and height

This list has S3 class PitchMemory2D.

Author(s)

Duncan Temple Lang

References

CUDA SDK e

See Also

cudaMallocPitch and cudaMemcpy2D

Examples

## Not run: 
  AB = matrix(rnorm(30 * 5), 30, 5)
  mem = mallocPitch( ncol(AB), nrow(AB), "float")
  mem[] = t(AB)

## End(Not run)

duncantl/RCUDA documentation built on May 15, 2019, 5:26 p.m.